---
name: claw_bot
description: claw-bot skill for Agent Harness decision loops over perception, action, and feedback MCP tools.
homepage: https://github.com/bozhi-ai/claw-bot
metadata: {"clawBot":{"skillKey":"claw_bot","environmentTypes":["realworld","simulator"]}}
---

# claw-bot Skill

Use this skill when an operator asks for a robot-operable task, realtime voice
control, environment perception, action planning, or execution feedback handling.

中文：当用户请求机器人执行任务、读取环境、处理语音指令、规划动作或根据反馈重试时，
使用 `claw_bot`。这个 skill 的目标是让决策 Agent 严格通过 MCP 能力清单工作。

## Core Workflow

1. Read the active environment and capability manifest with `robot_status`.
2. Use `robot_perceive` or `robot_subscribe` when the next decision depends on
   environment state.
3. Build the smallest safe action sequence that satisfies the operator intent.
4. Execute actions with `robot_act`, setting `environmentId` only when overriding
   the active environment.
5. Record controller or environment feedback with `robot_feedback`, then replan
   if feedback rejects or changes the action.

中文流程：

1. 先用 `robot_status` 看当前环境、runner、安全状态和能力 manifest。
2. 需要现场状态时调用 `robot_perceive` 或 `robot_subscribe`。
3. 生成最小、安全、可解释的 action sequence。
4. 用 `robot_act` 执行；只有需要覆盖 active environment 时才显式传
   `environmentId`。
5. 用 `robot_feedback` 记录控制器或环境反馈，并根据反馈继续规划。

## Hard Rules

- Treat `robot_*` MCP tools as the only perception and action interface.
- Do not use natural-language-only fallback for robot-operable requests when MCP
  tools are available.
- Do not assume any sensor, actuator, skill, policy, simulator, or robot model
  exists unless it appears in the capability manifest.
- Preserve explicit distances, directions, environment IDs, and operator
  constraints from the transcript.
- Use low-risk direct actions for simple stop, movement, posture, and turn
  requests.
- Use perception before action when the operator intent depends on current scene
  state.
- This repository is a claw-bot plugin only; it does not configure sensors,
  train skills, load policy networks, host simulators, or own low-level robot
  controllers. Robot Brain is the composition of this plugin and a decision
  agent.

中文硬规则：

- `realworld` 和 `simulator` 只是环境类型，不是工具名前缀。
- 不要猜测未在 manifest 中出现的传感器、执行器、技能或安全能力。
- 不要把语音转写当成直接动作 API；语音必须进入同一套决策循环。
- 简单动作走最小安全序列；依赖场景的动作先感知再执行。
- 反馈不是日志终点，必须能影响下一轮规划。

## Support Docs

- `README.md`
- `docs/overview.md`
- `docs/realtime-voice.md`